id stringlengths 3 3 | forecaster stringclasses 10
values | forecast_vintage stringlengths 7 95 | source_publication stringlengths 7 54 | geography stringclasses 2
values | horizon int64 2.02k 2.04k | unit stringclasses 3
values | metric stringclasses 6
values | scenario stringlengths 8 88 | shape stringclasses 4
values | value_basis stringclasses 3
values | low float64 106 1.2k ⌀ | high float64 106 1.07k ⌀ | point float64 106 1.07k ⌀ | base_year float64 2.02k 2.03k ⌀ | crypto_scope stringclasses 4
values | coverage stringclasses 5
values | state stringclasses 3
values | comparison_family stringclasses 4
values | superseded_by stringclasses 5
values | source_id stringlengths 3 3 | locator stringlengths 6 103 | qualification stringlengths 93 1.06k | method_description stringlengths 65 222 | model_reproduction stringclasses 1
value | maker_assessment stringclasses 1
value | support_review stringclasses 2
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
O01 | IEA | 2025 edition (source title; not stated in sheet cells) | 2025 edition (source title; not stated in sheet cells) | Global | 2,030 | TWh | annual_electricity | Base Case | point | table_value | 946 | 946 | 946 | null | unstated | total_facility_IT_plus_overhead | retained_historical | iea_global_base | O03 | S01 | Regional Data G57 (World), G54 Base Case, G55 2030, B56 total electricity consumption | Stored annex Base Case value for total, as distinct from IT-only, data-centre electricity consumption (sheet labels). The 2026 report's tabulated value for the same quantity is registered as O03. A one-unit difference between tabulated values is not interpreted as a substantive revision. The workbook's World Data sheet... | Bottom-up server-shipment model as described for the 2026 edition (S02 Annex A); the 2025 methodology is not retained; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O02 | IEA | 2025 edition (source title; not stated in sheet cells) | 2025 edition (source title; not stated in sheet cells) | US | 2,030 | TWh | annual_electricity | Base Case | point | table_value | 426 | 426 | 426 | null | unstated | total_facility_IT_plus_overhead | retained_historical | iea_us_base | O18 | S01 | Regional Data G59 (United States), G54 Base Case, G55 2030, B56 total electricity consumption | 2025-edition US Base Case projection for 2030 of total, as distinct from IT-only, consumption (sheet labels). The 2026 report restates the US 2030 Base Case total; that value is registered as O18. The workbook's World Data sheet distinguishes hyperscale, colocation/service-provider and enterprise facilities; the comple... | Bottom-up server-shipment model as described for the 2026 edition (S02 Annex A); the 2025 methodology is not retained; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O03 | IEA | 2026-04 (colophon month; release day not verified) | 2026-04 (colophon month; release day not verified) | Global | 2,030 | TWh | annual_electricity | Base Case | point | table_value | 945 | 945 | 945 | null | unstated | total_facility_IT_plus_overhead | refreshed | iea_global_base | null | S02 | Table A.1 printed p108 and Table A.4 printed p110 (World, 2030, Base Case); text at printed p10 and p24 | Tabulated Base Case value. The report text rounds this projection to about 950 TWh (printed p10 and p24) and describes it as close to the trajectory in the IEA's 2025 report. Model projection, not a measured outcome. | Bottom-up model driven by IT equipment shipments, covering servers, storage and network equipment inside facilities and excluding data transmission networks (Annex A methodology); published outputs read; model not rebuilt. | not_attempted | supported | supported |
O04 | LBNL | 2026-06-18 | 2026-06-18 | US | 2,028 | TWh | annual_electricity | Reference Case | point | text_value | 464 | 464 | 464 | null | excluded | servers_storage_network_infrastructure | retained_historical | null | null | S03 | Total Energy Use p8 | Reference Case lies within the old range according to the source. Interval inclusion does not prove an unchanged central forecast. | Bottom-up energy-use model; total-server shipment inputs from IDC's tracker through 2029 with 2030 extrapolated, and AI accelerator inputs from IDC compared with Omdia (section 3.1, printed p13); model not rebuilt. | not_attempted | supported | supported |
O05 | LBNL | 2026-06-18 | 2026-06-18 | US | 2,030 | TWh | annual_electricity | Reference Case with Compounded Uncertainty bounds | scenario_range | table_value | 521 | 843 | 649 | null | excluded | servers_storage_network_infrastructure | refreshed | null | null | S03 | Table 2 | 649 Reference Case; 521-843 Compounded Uncertainty scenarios. These are not a confidence interval. The Reference Case uses IDC data as-is, including the assumption that all global ASIC shipments go to the US (printed p15). | Bottom-up energy-use model; total-server shipment inputs from IDC's tracker through 2029 with 2030 extrapolated, and AI accelerator inputs from IDC compared with Omdia (section 3.1, printed p13); model not rebuilt. | not_attempted | supported | supported |
O06 | LBNL | 2026-06-18 | 2026-06-18 | US | 2,024 | TWh | historical_model_estimate | historical estimate | point | text_value | 192 | 192 | 192 | null | excluded | servers_storage_network_infrastructure | retained_historical | null | null | S03 | Total Energy Use p8 | Revised estimate of history, not metered national actual and not a forward-forecast revision. | Bottom-up energy-use model; total-server shipment inputs from IDC's tracker through 2029 with 2030 extrapolated, and AI accelerator inputs from IDC compared with Omdia (section 3.1, printed p13); model not rebuilt. | not_attempted | supported | supported |
O07 | EPRI | 2026 (white-paper citation year) | brief undated; cited white paper dated 2026 | US | 2,030 | TWh | annual_electricity | EPRI scenario range | scenario_range | text_value | 380 | 790 | null | null | included | small_large_data_centres_and_crypto | refreshed | null | null | S04 | PDF p2 annual-energy paragraph; PDF p3 Figure 2 caption | Approximate ('roughly') scenario bounds, not a point forecast. Brief inspected; the cited white paper is not retained. LBNL's comparison (S03) describes the EPRI range as roughly 390 to 800 TWh; the brief's own figures are registered. | Scenarios of project execution for operational, under-construction and planned sites, drawn from state-level capacity data; model not rebuilt. | not_attempted | supported | supported |
O08 | McKinsey | 2024-09 | 2024-09 | US | 2,030 | TWh | annual_electricity | medium scenario | point | chart_value | 606 | 606 | 606 | null | excluded_inferred | data_centres_unspecified | retained_historical | null | null | S05 | Exhibit 1 (medium scenario), 2030 column; footnote 1, p2 | Medium-scenario chart value. Footnote 1 states that the 2030 share calculation excludes cryptocurrency; applying that exclusion to this series is inference. The same exhibit is republished in the August 2025 McKinsey article (S06). | McKinsey analysis; Exhibit 1 cites Global Energy Perspective 2023 as a source; complete model not inspected. | not_attempted | supported | supported |
O09 | McKinsey | 2025-08 | 2025-08 | Global | 2,030 | GW | data_centre_capacity_demand | continued momentum | point | text_value | 220 | 220 | 220 | null | unstated | data_centres_unspecified | retained_historical | null | null | S06 | Opening and Exhibit 2 | Text states about 220 GW (Exhibit 2), and the exhibit labels 219 GW for 2030. Capacity demand, not installed capacity or annual consumption; no GW-to-TWh conversion. Exhibit 2 distinguishes AI and non-AI workloads; the complete facility and IT-versus-total-power boundaries remain unverified. | Published proprietary demand assessment; complete model not inspected. | not_attempted | supported | supported |
O10 | Goldman Sachs | 2025-02-04 | 2025-02-04 | Global | 2,030 | percent_growth | data_centre_power_demand_growth | unstated | upper_bound | text_value | null | 165 | null | 2,023 | unstated | data_centres_unspecified | retained_historical | gs_global_growth | O11 | S07 | Opening | As much as 165 per cent by the end of the decade versus 2023: an upper end, with no central point or lower bound stated. The article quantifies power demand in GW; no TWh basis is stated. Later registered vintages of this series are O19 and O11. The article discusses cloud, traditional and AI workloads and corporate an... | Public research summary; underlying forecast model not inspected. | not_attempted | supported | supported |
O11 | Goldman Sachs | 2026-03-04 | 2026-03-04 | Global | 2,030 | percent_growth | data_centre_power_demand_growth | unstated | point | text_value | 220 | 220 | 220 | 2,023 | unstated | data_centres_unspecified | retained_historical | gs_global_growth | null | S08 | Opening | 220 per cent versus 2023. Global geography is inferred: this page does not state it, and the September 2026 transcript (S09) places the March figure in a global discussion. The page names the immediately prior outlook value, registered as O19. Dated March 2026 by this primary summary, not April as in the legacy ledger. | Public research summary; underlying forecast model not inspected. | not_attempted | supported | partly supported |
O12 | Goldman Sachs | 2026-08-12 (statement recorded) | 2026-09-01 | Global | 2,030 | percent_growth | data_centre_power_demand_growth | unstated | point | text_value | 170 | 170 | 170 | 2,025 | unstated | data_centres_unspecified | refreshed | gs_global_growth | null | S09 | Transcript answer | About 170 per cent versus 2025, from a transcript that may differ from the audio. Do not compare 170 per cent (base 2025) directly with 220 per cent (base 2023). The same answer gives the March figure on the 2025 base, registered as O20. | Public research summary; underlying forecast model not inspected. | not_attempted | supported | supported |
O13 | BloombergNEF | 2025-12-01 | 2025-12-01 | US | 2,035 | GW | installed_capacity | unstated | point | text_value | 106 | 106 | 106 | null | unstated | data_centres_unspecified | retained_historical | bnef_us_installed | O14 | S10 | Opening sentence | The source words the value as data-centre power demand in a US context; its classification as installed capacity follows the later BNEF article and is inferred. That article names the December 2025 forecast as comparator for its 2030 installed-capacity figure; for 2035 the comparator is implied by date and the stated r... | Project-pipeline installed-capacity forecast; underlying model not inspected. | not_attempted | supported | partly supported |
O14 | BloombergNEF | 2026-07-21 | 2026-07-29 | US | 2,035 | GW | installed_capacity | base case | point | text_value | 194 | 194 | 194 | null | unstated | data_centres_unspecified | refreshed | bnef_us_installed | null | S11 | Item 1 | Capacity expected online, not measured output. The forecast release and the summary article dates differ. The 2035 sentence does not repeat the comparator date given for 2030 in the preceding sentence. The outlook distinguishes hyperscaler and non-hyperscaler capacity and discusses AI training, inference and colocation... | Project-pipeline installed-capacity forecast; underlying model not inspected. | not_attempted | supported | supported |
O15 | BloombergNEF | 2026-07-21 | 2026-07-29 | US | 2,030 | GW | installed_capacity | base case | point | text_value | 118 | 118 | 118 | null | unstated | data_centres_unspecified | refreshed | bnef_us_installed | null | S11 | Item 1 | Separate horizon from the 194 GW 2035 estimate. The outlook distinguishes hyperscaler and non-hyperscaler capacity and discusses AI training, inference and colocation; the complete equipment boundary is not specified. | Project-pipeline installed-capacity forecast; underlying model not inspected. | not_attempted | supported | supported |
O16 | S&P Global Energy 451 Research | 2026 edition | publication date unverified | US | 2,030 | TWh | annual_electricity | base forecast | point | text_value | 728 | 728 | 728 | null | unstated | data_centres_unspecified | refreshed | null | null | S12 | Opening summary | Primary publisher summary. Crypto and enterprise scope remain unverified; no scope claim is inherited from the old ledger. Goldman Sachs states that it uses a 451 Research data set (S09), so the two cannot be assumed independent; the exact input overlap is not established. | Public summary of the 451 Research base forecast; complete method not inspected. | not_attempted | supported | supported |
O17 | Deloitte | 2025 predictions edition | 2024-11-19 | Global | 2,030 | TWh | annual_electricity | conditional headline estimate | point | text_value | 1,065 | 1,065 | 1,065 | null | unstated | data_centres_unspecified | retained_historical | null | null | S13 | Opening paragraph and endnote 1; efficiency paragraph carrying reference 2 | Approximately 1,065 under stated assumptions. The paragraph carrying reference 2 also gives about 1,000 TWh with continued efficiency gains and above 1,300 TWh without them. The article cites Goldman Sachs and EPRI among its sources (endnote 1). The article discusses computing, storage, cooling and networking but does ... | Conditional estimates drawing on public sources and specialist conversations (endnote 1); no formal model documented or rebuilt. | not_attempted | supported | supported |
O18 | IEA | 2026-04 (colophon month; release day not verified) | 2026-04 (colophon month; release day not verified) | US | 2,030 | TWh | annual_electricity | Base Case | point | table_value | 426 | 426 | 426 | null | unstated | total_facility_IT_plus_overhead | added | iea_us_base | null | S02 | Table A.4 printed p110 (United States, total electricity consumption, 2030, Base Case) | United States total consumption row beneath North America in Table A.4. Model projection, not a measured outcome. | Bottom-up model driven by IT equipment shipments, covering servers, storage and network equipment inside facilities and excluding data transmission networks (Annex A methodology); published outputs read; model not rebuilt. | not_attempted | supported | supported |
O19 | Goldman Sachs | before 2026-03-04 (prior outlook; date not stated) | 2026-03-04 | Global | 2,030 | percent_growth | data_centre_power_demand_growth | unstated | point | text_value | 175 | 175 | 175 | 2,023 | unstated | data_centres_unspecified | added | gs_global_growth | O11 | S08 | Opening summary ('vs +175% previously') | The March 2026 summary states this as its immediately prior outlook. That prior outlook's own publication is not retained. Global geography is inferred through O11. | Public research summary; underlying forecast model not inspected. | not_attempted | supported | partly supported |
O20 | Goldman Sachs | 2026-03 (beginning of March, as stated in S09) | 2026-09-01 | Global | 2,030 | percent_growth | data_centre_power_demand_growth | unstated | point | text_value | 117 | 117 | 117 | 2,025 | unstated | data_centres_unspecified | added | gs_global_growth | O12 | S09 | Transcript answer ('We were at about 117% back at the beginning of March') | About 117 per cent versus 2025, stated in August 2026 for the March outlook; that this restates O11 on a different base is inference. Transcript may differ from the audio. | Public research summary; underlying forecast model not inspected. | not_attempted | supported | supported |
O21 | McKinsey | unresolved; the page cites a McKinsey article of 2025-12-17 (not retained) | 2026-06-30 | Global | 2,030 | GW | data_centre_capacity_demand | current adoption scenarios | point | text_value | 220 | 220 | 220 | null | unstated | data_centres_unspecified | added | null | null | S14 | Opening paragraph, sentence carrying endnote 1 | Text states that under current adoption scenarios global data-centre demand could almost triple between 2025 and 2030, from about 82 GW to about 220 GW. The sentence carries endnote 1, citing McKinsey's article of 17 December 2025, which is not retained; that is the cited article's publication date and the underlying f... | Not described beyond the cited December 2025 article; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O22 | McKinsey | 2026-07-31 | 2026-07-31 | US | 2,030 | GW | data_centre_capacity_demand | unstated; the page calls the trajectory relatively certain in all scenarios through 2030 | point | text_value | 121 | 121 | 121 | null | unstated | it_load | added | null | null | S15 | Section 'How might power demand play out in the next decade?', sentence citing Exhibit 1 | Text states that McKinsey's model suggests data-centre power demand could grow about 27 per cent a year through 2030, reaching 121 GW of data-centre IT demand. IT demand, not total facility power. The page assesses US data-centre power demand and calls the trajectory relatively certain in all scenarios through 2030. En... | Data-centre evolution scenarios and a state-by-state grid-capacity analysis, as described on the page; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O23 | McKinsey | 2026-09-18 | 2026-09-18 | Global | 2,035 | GW | data_centre_capacity | unstated | point | text_value | 389 | 389 | 389 | null | unstated | data_centres_unspecified | added | null | null | S16 | Section on rising power demand, global capacity sentence following the sentence carrying endnote 10 | Text states that global data-centre capacity is expected to grow about 17 per cent a year, reaching about 389 GW by 2035, a 4.6-fold increase. That sentence carries no endnote; endnote 10, citing McKinsey Data Center Solution, is attached to the preceding sentence on data-centre power demand. Worded as capacity; the re... | The capacity sentence has no endnote; endnote 10 cites McKinsey Data Center Solution for the preceding power-demand sentence; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O24 | Gartner | unresolved; the 2026-06-10 release cites Gartner's forecast report labelled 1Q26 (not retained) | 2026-06-10 | Global | 2,030 | GW | data_centre_capacity_demand | unstated | point | text_value | 290 | 290 | 290 | null | unstated | data_centres_unspecified | added | null | null | S17 | Opening paragraphs, sentence following the 2026 power-demand figure | Text states worldwide data-centre power demand of 132 GW in 2026, up from 104 GW in 2025, and estimates 290 GW by 2030. The release names its underlying report as 'Forecast: Data Center Power Capacity and Consumption, Worldwide, 2024-2030, 1Q26', which is not retained, so the forecast vintage is unresolved. The same re... | Not described in the press release; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O25 | S&P Global | 2026-06 (webinar report); underlying forecast vintage unstated | 2026-06-15 | US | 2,030 | GW | data_centre_capacity | unstated | point | text_value | 151.734 | 151.734 | 151.734 | null | unstated | data_centres_unspecified | added | null | null | S18 | Paragraph beginning 'The US dominates global data center capacity' | Text states that US capacity is projected to reach 151,734 MW by 2030, from 62,242 MW as of March 2026; recorded as 151.734 GW, an exact unit conversion. S&P Global publishes the article, which reports an S&P webinar featuring a 451 Research analyst; the projection is not attributed within its sentence, so 451 Research... | Not described in the article; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O26 | Deloitte | 2025-04-09 | 2025-04-09 | US | 2,035 | GW | data_centre_capacity_demand | Deloitte estimate | point | text_value | 176 | 176 | 176 | null | unstated | data_centres_unspecified | added | null | null | S20 | Opening section, sentence citing figure 1 | Text states that, according to Deloitte estimates, data-centre electricity demand could rise five-fold by 2035, reaching 176 GW. The sentence does not name a geography; US is inferred from the article's US context and from Deloitte's February 2026 article (S21), which restates 176 GW for data centres nationally in a US... | Deloitte Research Center analysis of named external data and studies (endnote 1); method not described; published outputs read; model not rebuilt. | not_attempted | supported | supported |
O27 | Gartner | unresolved; the 2026-06-10 release cites Gartner's forecast report labelled 1Q26 (not retained) | 2026-06-10 | Global | 2,030 | TWh | annual_electricity | unstated | lower_bound | text_value | 1,200 | null | null | null | unstated | data_centres_unspecified | added | null | null | S17 | 'Energy-Constrained Future', opening sentence | Text states that data-centre electricity consumption is estimated to reach over 1,200 TWh by 2030; recorded as an exclusive lower bound with no point or upper value. Figure 1 reports 2025 to 2027 consumption including cooling and other infrastructure; whether the 2030 statement uses the same boundary is not stated. An ... | Not described in the press release; the release cites a 1Q26 Gartner forecast report; published outputs read; model not rebuilt. | not_attempted | supported | supported |
AI Energy-Demand Forecast Scorecard
Data, code and paper for The AI Energy-Demand Forecast Scorecard: what data-centre electricity forecasts measure and what can be checked, by NM AI Research (ORCID 0009-0003-4213-7769). Licence: CC BY 4.0.
This work was produced through an AI-assisted workflow directed and reviewed by the author.
- Version 0.5: the folder
v0.5/, byte-identical to the Zenodo record https://doi.org/10.5281/zenodo.23434495 - Version 0.4: the folder
v0.4/, byte-identical to the Zenodo record https://doi.org/10.5281/zenodo.23384001 - All versions: https://doi.org/10.5281/zenodo.20572928 (resolves to the latest version)
- Code and interactive page: https://github.com/NMAIResearch/forecast-scorecard
Version 0.5
Forecasts of data-centre electricity demand from different publishers measure different quantities, draw different boundaries and date from different editions. Each registered figure is tied to a retained publication and an independent claim-support receipt, with dates, edition, target and base years, unit, quantity, scenario and coverage kept as separate fields. Candidate figures reach beyond the ledger's Global and US figures, to the Australian National Electricity Market, Europe and China as the IEA defines them, France, Denmark, Italy, Great Britain, Norway, Germany, Finland, Sweden, Belgium, Poland, Switzerland and the European Union. A revision is computed only where one publisher states both figures for the same quantity and both carry supported receipts, and every declined comparison is listed. The paper makes no demand forecast, accuracy ranking or capacity-uplift estimate.
Configurations
forecasts(default):v0.5/FORECASTS.csv, a generated view of the version 0.5 ledger with one row per registered observation, including its source, qualification and support review. The authoritative ledger isv0.5/OBSERVATIONS.json; the candidate observations and their receipts are inv0.5/CANDIDATE_OBSERVATIONS.jsonandv0.5/CANDIDATE_SUPPORT_REVIEW.json.forecasts_v0_4:v0.4/FORECASTS.csv, the same view of version 0.4.legacy_v0_3:v0.5/LEGACY.csv, the version 0.3 dataset, byte-preserved; its earlier verified labels are not current acceptance.interactive_v0_2:forecast_scorecard_data.csv, the 11 forecasts of the v0.2 dataset; its "verified" labels are not current claim-support verdicts.
Files
v0.5/: the 17 files of version 0.5. Start withPAPER.mdandREADME.mdthere.python3 -B reproduce.py --check, run from that folder, checks the package with Python 3.11 or newer, the standard library only and no network access.v0.4/: the 17 files of version 0.4, unchanged.forecast_scorecard_data.csvandbuild.py: the v0.2 data and the builder of its interactive page.LICENSE: Creative Commons Attribution 4.0 International.
AI assistance, conflicts and limitations
The statement for version 0.5 is in v0.5/README.md and at the end of the paper. Independent analysis and open-science documentation only, not investment advice.
Citation
NM AI Research. The AI Energy-Demand Forecast Scorecard: what data-centre electricity forecasts measure and what can be checked. Version 0.5. Zenodo. https://doi.org/10.5281/zenodo.23434495. Licensed CC BY 4.0.
- Downloads last month
- 87